OpenAI Privacy Filter

2026-04-22 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI releases Privacy Filter, a small open-weight model under the Apache 2.0 license that detects and masks personal data in text and can run locally.

What changes

  • Bidirectional token classifier (1.5 billion parameters, 50 million active) derived from an autoregressive model close to gpt-oss.
  • Detects eight categories: people, addresses, e-mails, phone numbers, URLs, dates, account numbers and secrets (passwords, API keys).
  • Can run on a workstation or in a browser, so unfiltered data is not sent to a server.
  • Can be tuned by fine-tuning: F1 rises from 54% to 96% on a domain adaptation benchmark.

### Specifications - Apache 2.0 license, weights on Hugging Face and GitHub. - Context: 128,000 tokens. - Modality: text, constrained Viterbi span decoding. - Identifier: openai/privacy-filter. ### Pricing Not applicable: open weights, to be self-hosted. ### Benchmarks | Evaluation | Score | | --- | --- | | PII-Masking-300k (F1) | 96% (precision 94.04%, recall 98.04%) | | PII-Masking-300k corrected for annotation errors (F1) | 97.43% | Scores published by the vendor. ### Key takeaways A ready-to-use compliance building block to anonymize logs, indexes and training sets without leaving the company perimeter.

Links

Tags : Open Source, OpenAI, Open-weight, Confidentialité, PII, Classification